Advances in molecular medicine are providing many new treatments that promise to be safer and more effective than traditional cytotoxic treatments by targeting the molecular characteristics of each patient's tumor (1–3). As these new targeted treatments enter clinical trials, there is a growing need to derive quantitative characteristics from images of cancer lesions (“quantitative imaging biomarkers”) that accurately assess the clinical benefit of these treatments (surrogate endpoints in clinical trials). Tumor shrinkage is the hallmark of response to traditional cytotoxic cancer therapies (4), and thus linear measurement of target lesions is the imaging biomarker used in most clinical trials using criteria such as Response Evaluation Criteria in Solid Tumors (RECIST) (5–7), Response Assessment in Neuro-Oncology (RANO) (8, 9), and International Harmonization Criteria (10). However, targeted, noncytotoxic therapies may arrest cancer growth and improve progression-free survival without necessarily shrinking tumors (11–14). Simple linear measurement may underestimate treatment response (15–18), in addition to having other limitations (7, 19). Alternative imaging biomarkers may be more promising than linear measurement for assessing response, especially with targeted therapeutic agents, as they can capture specific imaging features related to biological alterations in tumors during treatment (eg, heterogeneity, hypoxia, or changes in tumor microenvironment) (20–24), unlike tumor shrinkage (15, 25–27). Indeed, quantitative imaging biomarkers that reliably detect the results of anticancer agents (as opposed to detecting only change in tumor size) are desirable for all classes of therapeutic agents (28). Such new imaging biomarkers could become surrogate endpoints in clinical trials, as regulatory approval can be based on surrogate endpoints that document clinical benefit (29).
Development of imaging biomarkers follows a life cycle, starting with discovery and validation (“emerging biomarkers”), then translation and incorporation into clinical trials, and eventually to qualification for clinical use as surrogate endpoints for evaluating treatments (“qualified biomarkers”) (30). A number of research groups are working on the discovery/validation of the spectrum and developing new quantitative imaging biomarkers, including the Quantitative Imaging Network (QIN) (31) and the broader community (32–39). On the translation end of the spectrum, many of the new imaging biomarkers are ready to be translated for use in clinical trials, such as tumor volume (40), changes in contrast enhancement on computed tomography (41), radiotracer uptake on positron emission tomography (PET) (32, 42–46), kinetic parameters in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) (47–49), and spatial maps of such parameters (50, 51); however, very few of these new imaging biomarkers have yet to be incorporated into clinical trials for assessing treatment response.
Current image viewing and annotation tools are limited in their ability to support incorporating new imaging biomarkers into clinical trials in 4 major ways. First, although there are several commercial and open-source tools available to assess cancer lesions (52–55), they generally support very few measures of cancer lesions, such as linear dimension of target lesions, and they cannot be readily extended to deploy novel imaging biomarkers. Newer algorithms for computing imaging biomarkers are generally written in a variety of languages such as Java, Python, and C/C++, or exist within single toolkits [eg, MATLAB and 3D Slicer (56, 57)], which may not be compatible with current image assessment tools. Second, current lesion assessment tools are designed for only tracking cancer lesions in clinical practice, and they generally do not provide workflow management and study oversight features needed for assessing new image biomarkers in clinical trials. Third, there are no decision tools that use new imaging biomarkers for assessing treatment response in patients and overall drug effectiveness in clinical trial cohorts. Such decision-making requires calculating a variety of response measures in patients and across cohorts—tasks generally done by hand, making it difficult to compare multiple alternative imaging biomarkers. Fourth, it is costly and difficult to accrue aggregate data needed to qualify new imaging biomarkers as surrogate endpoints for clinical trials (58). Qualification of new imaging biomarkers requires collecting context-specific assessments of the performance of the biomarker relative to clinical outcomes (59). It is challenging to acquire sufficient data that link imaging biomarker data with clinical outcomes, such as survival (60). Efforts such as the Quantitative Imaging Biomarker Alliance (QIBA) are creating consensus on processes to qualify new imaging biomarkers (61), but their ultimate success depends on expanding public data sets (62) and leveraging many studies from individual laboratories and cooperative groups, which currently cannot be repurposed for this task because image annotations (or biomarker values extracted from them) are not recorded in standardized formats.
We developed ePAD—one of the research projects of the QIN—to address all of these challenges by developing a modular software platform integrating image viewing with computation of emerging and validated quantitative imaging biomarkers, facilitating translation of novel biomarkers into clinical trials as surrogate endpoints. In this paper, we will present ePAD's core architecture and describe the ways in which it meets the foregoing challenges. We also describe active research projects that are leveraging ePAD.
The ePAD Platform
We describe the design of ePAD and its core architecture, presenting this information from 4 different perspectives that address 4 major challenges mentioned above: (1) as a platform enabling the computing of novel imaging biomarkers of cancer treatment response, (2) as a workflow management and study oversight tool enabling the oversight for assessing new image biomarkers in clinical trials, (3) as a clinical decision support tool for the treatment response assessment using current and new imaging biomarkers, and (4) as infrastructure to permit researchers to aggregate evidence needed to show that new imaging biomarkers predict survival, which can be useful in qualifying them as surrogate endpoints in clinical trials.
Image Annotations in ePAD
A key distinguishing feature of ePAD is its support for standardized formats for image annotations, specifically Annotation and Image Markup (AIM) (63) and DICOM segmentation objects (64). AIM is an information model developed by the National Cancer Imaging Program of NCI for storing and sharing image metadata (65–67), such as lesion identification, location, size measurements, regions of interest (ROIs), radiologist observations, anatomic locations of abnormalities, calculations, inferences, and other qualitative and quantitative image features. The image metadata also include information about the image, such as the name of imaging procedure and how or when the image was acquired. AIM supports controlled terminologies, enabling semantic interoperability. In the use case of lesion annotation in cancer, the value of AIM is recording lesion identifiers (enabling unambiguous tracking of lesions across longitudinal images), anatomic locations of lesions, lesion types (target, nontarget, new lesion, or resolved lesion), and study types (baseline or follow up). This semantic information is critical for automating the generation of tabular summaries of lesions, and it also enables automating comparing the response assessment in patients according to different imaging biomarkers (see Section “Clinical Decision Support Tool for Treatment Response Assessment”). AIM has recently been incorporated into DICOM Structured Report (DICOM SR) (68), with specifications for saving AIM in DICOM-SR (69).
Architecture of ePAD
ePAD (70–72) is a freely available quantitative imaging informatics platform (http://epad.stanford.edu) distributed as a virtual machine or as Docker containers. Users can download virtual machine or Docker version of ePAD and host it in their own environment. This enables them to restrict the access to their private networks, typically to the hospital network. These machines generally do not have access to the internet. The core architecture of ePAD is shown in Figure 1. The ePAD platform comprises the following 5 main components: (1) the ePAD viewer, a zero-footprint web image viewer and image annotator, (2) ePAD web services, providing a programming interface to ePAD services, (3) an image database, (4) an image annotation database, and (5) plugin modules (server-side and client-side for extending the ePAD platform). The image database, image annotation database, and ePAD web services comprise the “back end” of ePAD. The ePAD plugin modules extend the functionality of ePAD, and while most of the plugins described below were developed by us, we also describe several developed by the community, and such community engagement will enable ePAD to foster an ecosystem enabling continued evolution of the platform to meet the needs of researchers broadly.
1. ePAD Viewer.
An important component of the ePAD viewer is its image annotation window (Figure 2). The ePAD viewer ensures the minimum information necessary to create a meaningful image annotation is collected from the user: the lesion name, the lesion type (target, nontarget, new lesion, or resolved lesion) and the anatomic location of the lesion, and the study time point (baseline or follow-up). The ePAD viewer automatically labels each lesion with a name (eg, “Lesion1”) to enable unambiguous determination of the same lesion on serial imaging studies (75). To specify the content of annotations, ePAD uses AIM templates (76) that are created by a separate freely available application. AIM templates specify the data elements to be provided by the user when making image annotations. All answer choices in ePAD templates are controlled terminology lists such as RadLex (77). The ePAD viewer prompts the user if certain values in the templates are inconsistent or incomplete (66). The ePAD viewer permits creating 2 types of ROI, coordinate based and pixel map based. The former is saved as coordinates in the AIM file (63), and the latter is saved as a DICOM segmentation object (64).
2. ePAD Web Services.
The ePAD viewer uses a set of RESTful Web services (78) to communicate with the back end of ePAD to retrieve images and save image annotations, as well as authenticating user credentials and invoking image calculation methods that need to be executed on the server. The ePAD Web Services provides programmatic access to the image database and the annotation database that are components of the ePAD back end (Figure 1). The ePAD Web Services is typically hosted on a server that resides within an institution's firewall so that all traffic between the ePAD viewer and the ePAD Web Services resides within the institution's Intranet. Thus, users can use ePAD to evaluate image data containing protected health information, provided the network on which ePAD is hosted is secure. Another model for hosting ePAD is a centralized, hosted version, which could provide publicly available images (which should be deidentified for public dissemination). The ePAD Web Services are used by plugin developers to extend ePAD's functionality, either as client-side or as server-side plugins (Figure 1). Plugin developers can use the ePAD Web services to access annotations and images in their own applications or to provide extensions to the ePAD platform.
3. Image Database.
Medical images in DICOM format are managed by an open-source PACS called dcm4chee (73). This PACS contains a DICOM image receiver and a programming interface that permits the ePAD Web Services to manage imaging studies within ePAD. The DICOM image database provides a temporary storage depot for images for image display and annotation in ePAD. The AIM annotations and DICOM segmentation objects in ePAD are saved indefinitely, however, as these annotations comprise the user-generated data in ePAD. Because DICOM images are large, the ePAD back end converts them into a lossless compressed PNG image object (“packed PNG”) that takes each 16-bit pixel in a DICOM image and packs it into a PNG color channel before returning it to the ePAD viewer, where it is unpacked. This approach significantly reduces the volume of data provided by the server and speeds performance of the ePAD viewer. To further speed image display performance, ePAD supports the Web-Accessible DICOM Objects [WADO (79)] protocol to retrieve lossy JPG images, while the lossless packed PNGs are initially loading.
4. Annotation Database.
As the user makes annotations on images in ePAD viewer, it creates AIM files. All AIM annotations are stored in an XML database [eXist (80)]. The AIM annotation database is accessible via functions in the ePAD Web Services, and it is the key resource that ePAD queries for lesion tracking and summarizing longitudinal changes in cancer treatment response, as described in Section “Clinical Decision Support Tool for Treatment Response Assessment.”
5. Plugin Modules.
Developers can create server-side and client-side plugins to access the data collected by ePAD to provide a new functionality. The server-side code can be written in a variety of languages, such as MATLAB, python, C/C++, or Java. We and other groups have created plugins to build a variety of features to address the challenges of (1) computing novel imaging biomarkers of cancer treatment response, (2) providing workflow management and study oversight features for assessing new image biomarkers, (3) creating clinical decision support tools for treatment response assessment using current and new imaging biomarkers, and (4) permitting researchers to aggregate evidence needed to show that new imaging biomarkers predict survival, which can be useful in qualifying them as surrogate endpoints in clinical trials.
Plugins currently available in ePAD are listed in the following sections.
JJVector Feature Extraction Plugin.
JJVector is a 2D feature extraction plugin we developed that analyses closed-shape annotations and extracts 2D radiomics features based on the intensity values from the ROI and the surrounding tissue of its associated organ (81). The plugin saves the calculated feature values in an AIM file that can be downloaded in different formats from ePAD, such as an excel summary sheet to be used in other applications such as training machine learning models.
ADLA Biomarker Plugin.
The Attenuation Distribution across the Long Axis (ADLA) plugin implements the ADLA semiquantitative imaging biomarker for assessing treatment response in solid malignancies and a measure of intralesional heterogeneity. We built this plugin in collaboration with prior works that created it (82, 83). ePAD calculates the standard deviation along the long axis to compute ADLA and saves it in the AIM file to be used for analyses such as response assessment as an alternative imaging biomarker. ePAD also generates an ADLA histogram of pixel values within the ROI when the long axis is selected (Figure 3).
Perfusion Analysis Plugin.
A contributor developed an ePAD plugin deploying an algorithm for computing T1 perfusion maps on dynamic contrast-enhanced studies based on his prior work (84). The plugin analyses the multiframe MRI images having different phases of dynamic contrast enhancement and calculates a T1 map for the imaged volume. The plugin scales the T1 map to 8 bits to save as a standard DICOM object (a probability DICOM Segmentation object) and paints the mask on the image using a color lookup table (Figure 4).
Riesz Texture Feature Plugin.
A contributor developed an ePAD plugin that computes image texture features based on Riesz wavelets (85). The latter are a subtype of convolutional approaches that can quantify image derivatives of any order and at multiple scales. The image derivatives are aligned along dominant local orientations, allowing characterization of the local organization of the image direction, with invariance to the local orientation of anatomical structures. These image derivatives have an intuitive interpretation, and the Riesz features have shown to provide valuable imaging measurements in various medical applications.
Quantitative Image Feature Engine (QIFE).
QIFE is an open-source feature-extraction framework we created that computes 3D radiomics features for ROIs that are created as DICOM segmentation objects (86). ePAD stores these image features in an AIM file for further analysis in radiomics studies or as alternative imaging biomarkers of response.
Quantitative Feature Explore (QFExplore) Plugin Suite.
The Quantitative Feature Explore (QFExplore) is a suite of plugins we developed for the ePAD platform, enabling the exploration and validation of imaging biomarkers in a clinical environment (85). Imaging features that can be extracted using QFExplore include histogram bins of Pixel Intensity Distributions (PID), statistical moments of PIDs (ie, mean, standard deviation, skewness, kurtosis), gray-level co-occurrence matrices (GLCMs), and Riesz wavelets (87). Figure 5 illustrates QFExplore plugin suite's feature comparison functions in action. The ROIs are visualized on the left, while color-coded gray-level co-occurrence matrices values are displayed in a chart on the right.
Quantitative Feature Pipeline (QIFP).
We created the QIFP, a cloud-based platform for building processing pipelines of image analysis algorithms (88). It provides a Docker library of image analysis algorithms for preprocessing, segmentation, and feature extraction that can be assembled into pipelines. The QIFP is integrated with ePAD so that any processing pipeline for generating quantitative imaging biomarkers can be executed in ePAD (or ePAD annotations can be consumed and used in QIFP processing pipelines).
ePAD includes several applications that are part of the platform and accessible via menu tabs in the ePAD viewer.
Computing and Comparing Imaging Biomarkers
A need that is critical for research is its ability to compute a variety of alternative imaging biomarkers besides linear dimension (used in RECIST and similar criteria). In a given clinical trial, patient response to treatment can be computed using a variety of imaging biomarkers, and a sizeable collection of data can be amassed if this is done across clinical trials that could ultimately be useful in comparing and evaluating alternative imaging biomarkers as secondary endpoints of response. Different imaging biomarker algorithms are written in different languages, and ePAD enables incorporating them into its image analysis workflow through its plugin mechanism described above. These plugins can execute source code modules written in MATLAB, Java, C/C++, or other languages, letting biomarker algorithm developers add their existing code to ePAD easily.
When users make annotations on images, ePAD automatically analyzes each annotation to generate the image biomarkers that the user chooses, and it saves them in AIM format. It also computes the minimum, maximum, standard deviation and mean for all the pixels that are inside the ROI. If the ROI is a line, ePAD calculates the length. If the ROI comprises 2 perpendicular lines, ePAD will calculate the length of the long axis and short axis. Additional features and biomarker candidates can be calculated by various plugins.
Workflow Management and Study Oversight
The ePAD viewer includes an application that provides a summary panel of annotations designed to streamline the task of summarizing for the radiologist all prior measurements and images in prior studies of each patient to convey the list of lesions previously measured, and which need to be measured on the current study. To populate this summary display, the ePAD viewer queries ePAD's annotation database to find all the lesions from the prior exams and list them for the user. This provides the user with a worklist of lesion measurements that need to be made for each imaging study. It also links each measurement to the image from which it was obtained. When the user clicks on a measurement, the corresponding image is retrieved and the measurement is displayed.
ePAD also facilitates oversight and managing image readings for clinical trial researchers and study administrators via user roles, worklists, and study progress monitoring. Project owners and administrators can create users and assign them specific roles to control their access to imaging data and annotations created by other users. Users or study supervisors can create worklists for people and assign to a reader. Using worklists allows the supervisors to divide the readings to multiple readers. A study progress monitoring application module in ePAD monitors the status of image annotations made in clinical trials and summarizes them in a table in the ePAD viewer. Study administrators can follow the image annotations made in multiple studies by group of users assigned to a particular study. The application can also track the progress of the annotation process by identifying which subjects/studies are fully annotated by all the annotators, which annotators have completed the annotation process for each subject and which subjects/studies have not yet been annotated yet (Figure 6). This functionality has been helpful the MGH/HST Martinos Center for Biomedical Imaging used this for MEDICI project (89), which used ePAD.
Clinical Decision Support Tool for Treatment Response Assessment
ePAD has applications to assist decision-making based on image biomarker assessments in the following 2 major cancer research tasks: determine treatment response in patients (ePAD longitudinal annotation report) and evaluate treatment effectiveness by determining the cohort-based treatment response (ePAD waterfall plot). We built these applications using ePAD Web services to retrieve AIM annotations and their associated images to track target lesions and compute cancer treatment response according to selected imaging biomarkers.
Longitudinal Annotation Reporting.
ePAD supports longitudinal annotation tracking, which provides a summary of quantitative image features across time. This is the basis for RECIST and other reports of response assessment. However, ePAD can generate such reports bases on any quantitative imaging biomarker it can collect from image annotations. It analyzes all the annotations of a subject and populates 3 dropdown menus to facilitate selecting them by shape, template, and measurement type (Figure 7). Users can select the basis for the longitudinal annotation report based on the selected measurement types. If a measurement is not present for a particular time point of a lesion, the table display it as a missing value. The summary section of the report will be filled automatically for the measurement type.
ePAD can generate a RECIST report by querying the annotations that are of linear type (Figure 7) and calculating sum of lesion dimensions on the images of each time point. RECIST report generation supports line and perpendicular lines annotations, as well as an image-based response rate (the percentage change in the sum of lesion dimensions compared with baseline). ePAD applies the RECIST rules to classify the response rate to determine the response category (ie, stable disease, partial response, complete response, and progressive disease). This information is displayed with the lesion measurements in the ePAD viewer (Figure 7). ePAD also checks the consistency of the annotations to determine if the anatomic location of the lesion is specified consistently on different time points of the patient; otherwise, the measurement will be marked as error to notify the user. The report also marks missing annotations for a lesion as error. The user can open the annotation in the ePAD viewer by clicking on the annotation measurement on a specific time point. The user can also open all annotations of the lesions on all time points by clicking the lesion name.
Besides using longitudinal measurement of lesions, ePAD can generate reports of lesion response based on other imaging biomarkers, such as the ADLA biomarker (82). The report evaluates the progress of the disease using the sum of ADLA scores for each timeline, similar to RECIST.
Waterfall plots are bar graphs showing the response of a cohort of patients to the same cancer treatment. The height of the bars represents the best overall response the patient had during the course of treatment, and each bar (patient) is ordered from best to worst response, which resembles a waterfall. These plots are highly useful for seeing how well a patient cohort responded to treatment, with the percentage of patients with positive response indicating effective treatment. ePAD generates waterfall plots of user-specified patient cohorts by computing the longitudinal annotation-based response in each patient in the cohort and ordering the response from best to worst response (Figure 8). The plot can be based on longitudinal measurement of lesions as the basis of evaluating response (ie, RECIST), but importantly, it can also be based on using newer imaging biomarkers of response such as ADLA or other imaging biomarkers that have been recently introduced by researchers. If the user selects to use RECIST, the waterfall plot module analyzes every subject in the cohort, generates the RECIST tables, gets the best response for each subject, and plots it in a decreasing order forming a waterfall plot. If the user selects to use ADLA, a waterfall plot is generated based on an ADLA table that ePAD computes for each subject, using the standard deviation of the line annotations on lesions as the measurement type (82). Then, the best response from the ADLA table for each patient is used to create the waterfall plot. Users can drill down to more granular data within the waterfall plot; the user can access the table that is used to make the best response rate computation by clicking the specific bar in the waterfall plot.
Application for Aggregating Evidence for Evaluating New Imaging Biomarkers
As noted earlier, ePAD has plugins to compute a variety of image biomarkers. Some of these plugins assume that cancer lesions are circumscribed, and if the images input into these biomarker plugins were annotated using only line annotations (eg, as part of RECIST measurements), ePAD can generate ROIs that circumscribe lesions automatically by executing image segmentation plugins that use the line or point annotations as seeds. In addition to segmentation, there are quantitative image analysis plugins that may operate on the entire image, and ePAD supports those as well. Current automated segmentation plugins and other analysis plugins available in ePAD are listed in the following sections.
Automated Segmentation in PET Images.
The plugin invokes automated segmentation of cancer lesions seen on PET images (90). It is triggered with a seed point ROI within the lesion. It analyzes the image volume to create a 3D ROI enclosing the lesion and creates a DICOM segmentation object marking the volume of the lesion. The DICOM segmentation object is added to ePAD with its associated AIM annotation file and displayed on the image series as a mask.
Automated 2D Lesion Segmentation.
ePAD has a 2D lesion segmentation plugin, LesionSeg. The plugin is triggered with drawing a polygon or a long axis line within a lesion. It analyzes the image and creates a polygon ROI stored as an AIM file containing the contours of the lesion (91).
Automated Image Segmentation of QF Explore Plugin Suite.
The QFExplore plugin suite has a plugin for automatically segmenting lungs in a DICOM image volume (85). The plugin analyses the volume, segments the lung volume, and creates a DICOM segmentation object.
To track usage statistics, ePAD collects anonymous data from all ePAD machines that are connected to internet (if the statistics are not disabled by the user). The statistics consist the number of users, projects, patients, studies, series, AIM annotations, DICOM segmentation objects, plugins, and templates that exist on the ePAD instances. Figure 9 shows the ePAD usage statistics collected from 2015 to 2018. For plugins and templates, the maximum number of entities is reported, as many are the same versions of the plugins and templates across ePAD instances. For all the other entities, the values reported by each ePAD instance are computed by getting the latest reported values for each year and summing them to obtain the total number. For example, in 2018, over 19,000 imaging studies were hosted in various ePAD instances worldwide, and over 55,000 annotations were created in ePAD on those imaging studies. As ePAD collects only the number of entities for privacy purposes, the numbers are cumulative; that is, this does not mean 55,000 annotations were created during 2018, but it means that at the end of 2018, 55,000 annotations existed on ePAD instances. In addition, currently there are a maximum of 11 plugins and 35 templates that are being used across all ePAD instances.
One of the key aims of ePAD is to facilitate collaborations among research sites and repurposing of their existing data, which we achieve by supporting standards and interoperability for images and annotations.
ePAD saves all image annotations that it collects using existing standards, in particular AIM (63) and DICOM segmentation objects (64), for volumetric ROIs. ePAD also supports the DICOM-SR standard via the dcmqi library (92) for volumetric ROI annotations. Recently AIM was harmonized with the DICOM standard, which provided DICOM-SR support of AIM annotation types under Supplement 200 with specifications for saving AIM in DICOM-SR (68, 69). ePAD also supports DICOM radiation therapy (DICOM-RTs) and tiff image files. ePAD analyses the DICOM-RT objects and extracts its ROI contours using the DICOM file interface library developed by MAASTRO (93). It then creates a DICOM segmentation object for each contour and saves it and an AIM file. ePAD also supports uploading tiff files and creates a DICOM image series from them using the patient identification number, patient name, study description, and series description supplied by the user. The file list is analyzed, and a DICOM file is created for each tiff file. The instance numbers of the DICOM files are ordered in the alphabetical order of tiff filenames.
Use Cases for ePAD in QIN and Other Research
Many research studies in that require viewing and annotating radiology images for making measurements of lesions or extracting radiomics features from them could benefit from using ePAD. Clinical trials of cancer treatments can be particularly helped given ePAD's workflow support and multireader support features, its support of interoperability standards, as well as its ability to compute many imaging biomarkers seamlessly within routine image annotation workflow. ePAD has been used by many researchers worldwide to support clinical research and clinical trials, and it has supported many published studies (75, 81, 85, 95–110), and it has been shown to improve the workflow of measuring target lesions (111). We briefly highlight support it has provided several projects in NCI's QIN.
In collaboration with Vanderbilt QIN, “Quantitative MRI for Predicting Response of Breast Cancer to Neoadjuvant Therapy” in which this group developed algorithms for computing quantitative perfusion maps of MRI images to deduce biomarkers of treatment response (112), we deployed their biomarker algorithms as a plugin to ePAD. As these researchers incorporate these perfusion analyses into clinical trials, ePAD will be able to deploy them as part of the image interpretation workflow.
Dana Farber Cancer Institute QIN.
The QIN project at Dana Farber, “Genotype and Imaging Phenotype Biomarkers in Lung Cancer (113),” developed pyRadiomics, a flexible platform that extracts a large panel of predefined features from medical images and is useful in characterizing cancer lesions. We incorporated pyRadiomics into ePAD as a Docker module that runs on the QIFP platform (88) (see above) so that users can invoke generation of these image features as part of image analysis workflows in clinical trials.
The QIN project within the ECOG-ACRIN cooperative group, “ECOG-ACRIN-Based QIN Resource for Advancing Quantitative Cancer Imaging in Clinical Trials,” is leveraging ePAD as a testbed for evaluating the deployment of imaging biomarkers into clinical trials. Currently this project is comparing ability of ePAD to evaluate a variety of quantitative imaging biomarkers as part of the routine workflow of image viewing and annotation in clinical trials.
American College of Radiology (ACR) Core Laboratory.
The ACR has a data archive and research toolkit called DART Portal (114) that operates as a gateway to browse and query data for research, quality improvement, and clinical study operational purposes. They are adding ePAD as an interface to DART to enable collecting image annotations as part of clinical trials in AIM format and storing that in DART.
Response assessment in patients with cancer in clinical trials is based on analysis of CT and magnetic resonance images (115). Objective criteria, such as RECIST, are critical to evaluation of response assessment in clinical trials, but lesion measurements vary with user experience, and they are often inconsistent or incomplete (105). There is a pressing need to recognize signals in radiology images that optimally assess and predict response to treatment. Tumor shrinkage is the hallmark of response to cytotoxic cancer therapies (4), and thus, linear measurement of target cancer lesions is the imaging biomarker used in current response criteria such as RECIST and International Harmonization Criteria (10). However, new targeted, noncytotoxic therapies arrest cancer growth and improve progression-free survival without necessarily shrinking tumors (11–14); thus, simple linear measurement may underestimate treatment response (15–18), and may not be the best proxy for tumor activity. To address these limitations, researchers are developing quantitative imaging biomarkers that may better assess the benefit of new treatments, but they have been challenging to introduce into clinical trial workflows.
In the paper, we have presented ePAD from 4 different viewpoints to highlight how it addresses key challenges for incorporating quantitative imaging biomarkers into clinical trials. First, it provides a platform for computing a variety of imaging biomarkers. Second, it provides a workflow management and study oversight tool enabling oversight for assessing new image biomarkers in clinical trials. Third, it provides clinical decision support tools to help clinical researchers assess treatment response using current and new imaging biomarkers. Fourth, ePAD provides infrastructure to permit researchers to aggregate evidence about how well imaging biomarkers predict response, which may help in qualifying them as surrogate endpoints in clinical trials.
There are many existing commercial and freely available tools available for medical image viewing and annotation, and although ePAD provides similar capability in terms of image viewing and drawing shapes on images, it provides many unique features that address many unmet needs in evaluating images clinical research. Osirix (94) and ClearCanvas (55) are 2 medical image annotation applications that provide similar image viewing capabilities, although they depend on a thick client, limiting collaboration as they are platform-dependent, while ePAD is a web-based viewer and requires no installation for users other than hosting a single instance of the ePAD machine for all users. In addition, Osirix saves its annotations in propriety format and supports exporting ROIs using ExportROIs plugin. On the other hand, ePAD and ClearCanvas supports AIM format. ePAD also supports the new DICOM-SR AIM object. Beyond the open-source tools, we recognize that several commercial tools are available for image viewing and analysis to enable response assessment. These tools were developed to enable evaluating established criteria such as RECIST in clinical trials; however, such tools are not optimal for research studies that wish to include novel imaging biomarkers of treatment response (eg, those being developed by NCI's QIN). This gap was a primary motivator for developing the ePAD system.
3D Slicer (56, 116), ImageJ (117), and MIPAV (118) are additional freely available desktop applications. 3D Slicer is a cross-platform open-source software for visualization and image computing. It has a plugin architecture to enable researchers to develop their algorithms via C++ plugins and Python scripted modules. It supports DICOM standard for the volumetric annotations and DICOM Structured Report (DICOM-SR) for the measurements collected from the ROIs. ImageJ and MIPAV are both Java applications that can run on any Java-enabled operating system, and researchers can develop their own plugins using Java language. Imagej2, an extended version of ImageJ, supports writing plugin scripts in various programming languages. ImageJ saves the labels and annotations as modified versions of the images or propriety ROI file formats, and MIPAV uses an Extensible Markup Language (XML) format they introduced in an effort to make their format readable by researchers. Although all 3 applications are cross-platform, they are desktop applications for a single user, which makes multiuser collaboration more difficult.
ePAD was developed to facilitate collecting annotations and measurements on target lesions in compliance with standards in the cancer imaging community. ePAD makes sharing code, data and annotations easy being a web application and saving the collected annotation data in well-documented and standardized formats [DICOM segmentation objects (64) and AIM (63) in particular].
In addition to providing standards-based storage of annotations, ePAD enables user-defined templates for flexible capture of information in the form of data collection templates as part of the annotations. The ePAD platform is also extensible via plugins that lets researchers implement analysis codes as server-side modules in MATLAB or other languages. Many plugins for segmentation and quantitative image biomarker computation are included with ePAD, and users can add additional biomarker modules.
Other functionalities of ePAD that differentiates it from similar existing image viewing applications is that ePAD supports important features unique to image analysis in clinical trial workflow. Specifically, ePAD provides tools enabling oversight of annotations as part of clinical trials, and it lets the users create a collaborative environment by creating projects and assigning users appropriate rights to limit their access facilitating large studies with multiple annotators. ePAD also provides decision support tools—longitudinal annotation summary and waterfall plots—that help researchers evaluate individual patient and cohort population treatment response, respectively. Finally, by computing a variety of image biomarkers on cohorts of patients, ePAD can accumulate a substantial amount of data that can permit studies comparing effectiveness of different imaging biomarkers as indicators of treatment response.
An ultimate metric of the success of ePAD will be increased use of the newer imaging biomarkers in clinical trials. This will require clinical trial groups to include computation of the biomarkers into their study protocols. As the community becomes aware of the potential of these methods and of the facility of tools such as ePAD to include them in clinical trials, we expect these methods will be more commonly used. Certainly, the amount of research studies undertaken to date using ePAD suggests promising future directions.